Working with AI · 6 min read
I’ve been using Claude for my business since 2023. I can usually spot AI-assisted writing in the first two sentences. Here are the patterns I look for, including the ones I still catch in my own drafts.
NA
Natalie AndersonMarketing consultant · Farmingdale, Maine
Every client project and every piece of content runs through AI at some point. And for most of that time, I’ve also been editing AI output for other people. Client drafts, marketing copy that someone fed through ChatGPT and sent over for review.
The patterns are incredibly consistent. Every model does the same things. And once you see them, you can’t unsee them.
AI writes competently. That’s almost worse than writing badly. The output looks professional enough that people hit publish without a second read. But it reads like every other AI-assisted post on the platform, and your audience can feel it even if they can’t name what’s off.
One thing before the list. This isn’t about hiding that you use AI. I use it for almost everything and say so. Nobody’s handing out prizes for typing every word yourself. These patterns are perfectly fine on their own, once in a while, but AI overuses them. Some people do want to hide their AI use. That’s not my concern. Overused, these patterns just produce bad writing. Flat, interchangeable, forgettable. Fix them because your writing gets better, not because you’re worried about getting caught.
1.AI overuses em dashes in short-form writing
I need to be honest about this one, because it’s personal. I banned em dashes from my own social media and short-form copy about two years ago. Em dashes are great punctuation. In a long article or a book, they do something a comma can’t.
The problem is that AI puts them in everything. A LinkedIn comment. A quick email. A Slack message. A two-sentence reply. Places where a period or comma would do the job without drawing attention to itself.
Before
We saw the trend — we jumped in — we scaled fast.
After
We saw the trend. We jumped in. Then we scaled.
The fix
Save em dashes for writing that earns them. For short-form content, replace them with periods or commas. If the sentence doesn’t work with simpler punctuation, it probably needs rewriting anyway.
2.The stack of three shows up in almost every AI draft
This is the one that will not die. AI groups things in threes more consistently than any other pattern, and it’s the hardest one to train out.
I have it in my project instructions. I have a dedicated file that says, explicitly, do not use lists of three. And it still happens. Regularly.
Before
Bold, innovative, and ready for the future.
After
Bold and built for what’s next.
The fix
Read your AI output out loud. Stacked threes have a cadence that’s easy to hear when you’re listening for it. The rule of three is fine when you’re choosing it. The problem is when your AI is choosing it for you on every other line.
3.Mirror constructions fake insight with contrast
“Not X. It’s Y.” AI loves this move. It’s confident and creates the illusion of insight by setting up a contrast. The first sentence establishes what something isn’t. The second delivers the revelation.
The problem is that mirror constructions almost always stand in for actual thinking. When you strip away the rhythm, most of them are saying something obvious.
Before
It’s not about the strategy. It’s about the execution.
After
Execution matters more than strategy.
The fix
Collapse it into a single direct statement. If the collapsed version sounds boring, the original probably wasn’t saying much either.
4.Throat-clearing openers announce a point instead of making it
“Here’s the thing.” “It’s worth noting that.” “What many people don’t realize is.” These are the filler phrases AI uses to warm up before saying something. They signal that a point is coming without actually making the point.
The related pattern is what I call the announcement colon. The sentence before the colon exists only to set up the sentence after it. Delete everything before the colon and the writing gets better every time.
Before
Here’s what I learned: the client didn’t want a strategy deck.
After
The client didn’t want a strategy deck.
The fix
Search your draft for “Here’s,” “The reality is,” and “It’s worth noting.” Delete the setup sentence and start with the actual point. For announcement colons, remove everything before the colon. If the remaining sentence stands on its own, and it almost always does, the setup was unnecessary.
5.The tidy closer ends every piece with a platitude
AI loves to wrap things up. A summarizing sentence that restates the theme. A philosophical bow that ties everything together with a little too much elegance.
The tidy closer is especially dangerous because it feels like good writing. It has warmth and rhythm. But it tells your reader that the ideas in this piece are small enough to be summarized in one platitude.
Before
At the end of the day, it’s not about the tools. It’s about the people behind them.
After
Nothing. The piece ends on its last real point.
The fix
Write your last paragraph, then delete the final sentence. See if the piece ends better without it. If you’ve made your point, you don’t need to restate it. If the last sentence is a question, delete it immediately. That’s almost always AI.
Why AI writing tells persist even with clear instructions
These patterns don’t go away because you know about them. I’ve spent two years building systems to prevent them. I have project-level instructions. I have reference files listing specific patterns to avoid. And I still catch at least one tell in almost every draft.
A marketing director at a major consumer brand posted this recently, in a piece about getting better results from AI:
Specimen · 12 words · 3 tells
“Here’s the thing: that’s not a Claude problem. That’s a depth problem.”
A throat-clearing opener, an announcement colon, and a mirror construction. Three of the five patterns in this article, stacked in two sentences, published by someone whose job is knowing how words land.
I’m not saying this to call anyone out. Anyone using AI can be guilty of it. For people deeply enmeshed in this space, the patterns become almost invisible from sheer exposure. And for anyone newer to working with AI, you’re so grateful for the speed that you don’t stop to notice how bland the writing is.
Every one of these patterns is a legitimate rhetorical device. That’s why the models learned them. The problem is frequency and context, not the devices themselves.
Setting up your AI to avoid these patterns is necessary. Reading your own output with enough skepticism to catch what slipped through is where the work actually happens. Every time. Not just when you remember to.
5 ways to catch AI writing tells in your own drafts
1
Add a "don't do this" file to your AI setup
Load a document listing specific patterns to avoid. Not “sound natural.” Specific patterns with examples. I built the one I use in my own projects.
2
Read your output out loud before you post it
AI tells have a rhythm. If a sentence sounds like it belongs in a TED talk, it’s probably a tell.
3
Delete your last sentence
Try it on your next draft. If the piece reads better without it, the closer was padding.
4
Search for "Here's" in your draft
The quickest way to find throat-clearing openers and announcement colons. Delete the setup, keep the point.
5
Ask your AI to review its own output for tells
Name the specific patterns rather than asking “does this sound like AI.” Adds two minutes, catches most of what a first read misses.
One more thing
Before publishing this, I ran it through Claude for editing suggestions. Tighten this paragraph, smooth that transition.
The edits came back full of mirror constructions, stacked threes, and announcement colons. Every pattern this article warns you about, introduced through the AI’s own suggestions. I spotted them within the first few paragraphs. Then I ran the tell check I just described and pulled out the rest.
I know these patterns. I teach people to spot them. And they still sneak into my work through the editing layer when I’m not paying attention. Knowing about them is step one. Catching them is the actual work, and it doesn’t get optional just because you’ve been doing this a while.
Free reference guide
The AI Writing Tells Checklist
All 14 patterns with before-and-after examples and fixes. Drop it into your Claude project or custom instructions so your AI knows what to avoid. Subscribers also get new posts on working with AI as they publish.
One useful email at a time. Unsubscribe anytime.
Tools I Recommend
Every one tested and vetted. Not all were used for this specific article.
Claude
My primary AI tool since 2023. Projects hold the voice rules and reference files that make this workflow possible.
Riverside
Remote recording that captures each side locally. Solves the audio quality problem before editing starts.
Some links are affiliate links. I only list tools I pay for and use in client work. It costs you nothing and does not affect what I recommend.
Frequently asked questions
What are the most common AI writing patterns to look for in your own drafts?
The most reliable signals are structural rather than word-level. Watch for mirror constructions, groups of three, em dashes in short-form writing, throat-clearing openers, and summarizing closers that restate the theme. Any one of these can appear in writing you did entirely yourself. Several stacked in a short piece usually means your AI made the choices.
Are em dashes a sign of AI writing?
Not on their own. Plenty of writers use em dashes well, particularly in long-form writing where they do work a comma can’t. The tell is frequency and context. Multiple em dashes in a short LinkedIn post or a quick email suggests AI, because the tool applies the same dramatic punctuation to every format regardless of whether it fits.
Do AI writing patterns hurt your SEO or AI search visibility?
Formulaic writing isn’t penalized directly, but it competes poorly. Search systems and answer engines favor content with specific claims and original examples. AI defaults toward generic phrasing, which produces content that reads like everything else on the topic and gives no reason to cite it over another source.
Can you stop AI from using these patterns with better prompts?
Only partly. Adding constraints to your project instructions or system prompt reduces the frequency considerably. It doesn’t eliminate it. These patterns are baked deep into how language models construct sentences, which means human review stays part of the process no matter how well you set things up.
What's the fastest way to edit AI tells out of a draft?
Paste the draft back into your AI and ask it to flag specific named patterns rather than asking a general question like “does this sound like AI.” Naming the patterns gives it something concrete to check against. This adds about two minutes and catches most of what a first read misses.